The shift
For the last few years, the first question executives asked about AI was which model to use. That question matters less every quarter. Capable models are available from several vendors, open-weight alternatives follow close behind, and prices keep falling. The model your organisation chooses this year will probably be replaced within two.
At the same time, AI is moving from answering to acting. The first systems drafted text and answered questions, and a person decided what to do with the result. The systems being built now place orders, change records, write to customers and call other software. They act on the organisation's behalf, often several steps at a time.
Taken together, these two changes move the source of advantage. When the model is interchangeable and the system acts, the outcome depends on the architecture around the model: what the system is allowed to touch, who approves what, how it is tested and monitored, what it costs to run, and who can change it. That architecture decides whether AI works, stays under control and pays off.
What it costs when the architecture is missing
Few organisations set out to build an AI architecture. They start with a use case, a vendor and a deadline, and the architecture becomes whatever the first projects leave behind. The costs arrive later, in four places.
Control over what agents can do
An agent that can act needs limits: which systems and data it can reach, which actions need a person's approval, and a record of what it did. Without them, its authority is whatever its credentials happen to allow. Gartner groups its top strategic technology trends for 2027 into three themes, and one of them is to govern the autonomous agent platform (Gartner, October 2026).
Systems the organisation cannot change alone
When the design, the code and the knowledge of how a system works sit with a supplier, every change becomes a new contract. Gartner predicts that by 2028, 70% of enterprises will abandon agentic AI built through vendor forward-deployed engineering, because costs soar and they cannot evolve the systems themselves. Its advice is to contract for knowledge transfer, intellectual property and exit (Gartner, September 2026).
Cost and value nobody can see
AI spending is spread across licences, cloud bills, supplier fees and staff time, and its returns are rarely measured in the terms a finance team uses. In McKinsey's 2026 survey, 37% of respondents attributed any EBIT impact to AI (McKinsey, August 2026). In our experience, part of the reason is that the measurement was never built into the systems in the first place.
Compliance you cannot prove
Rules for AI differ by jurisdiction and keep changing. Gartner predicts that fragmented AI regulation will extend to 75% of the world's economies by 2030 (Gartner, February 2026). An organisation that operates or sells across borders will be asked to show, system by system, what its AI does and how it is controlled. If that evidence has to be reconstructed after the fact, every new market and every audit becomes a project of its own.
A hypothetical example
Consider a company that adds an AI assistant to its customer service. This is an illustration, not a client story. The first version runs on one vendor's platform, answers questions and drafts replies. It works, so the assistant is given access to refunds and account changes. A second team builds a sales assistant on the same platform, and a third connects it to the order system.
A year later, nobody can say exactly what each agent is allowed to do, the AI spend sits in three budgets, and moving to a cheaper model means rebuilding prompts, integrations and tests that the vendor's team wrote. None of these steps was a bad decision on its own. Together they produced a system the company depends on but cannot see into or change alone.
When one vendor's stack is enough
Not every use of AI needs its own architecture. When the use is narrow, the stakes are low and the horizon is short, a single vendor's stack is often the sensible choice: a drafting assistant for one team, a summary tool for internal documents, a time-boxed experiment. The trade-off changes when AI starts to act on customers, money or core systems, when several teams build on it, or when the system is meant to outlast the current generation of models.
From and to
In many organisations today, AI is bolted onto one vendor's stack. Agents have unclear authority, running costs sit in other budgets, and the systems cannot be changed without outside help.
The direction we argue for is AI run as an owned system:
- one architecture that every team builds on;
- explicit boundaries for what each agent can do, enforced by the system itself;
- cost and value measured from the first release;
- models that can be swapped when a better or cheaper one appears;
- a team inside the organisation that can run and change it.
What good looks like
Under control
AI acts only within limits you set. Each agent has a defined scope: the data it can read, the actions it can take, and the points where a person must approve. Those limits are enforced by the system, so they still hold when a prompt or a model changes. Every action leaves a record you can show to a customer, an auditor or a regulator. Requirements such as the EU AI Act, ISO/IEC 42001 and the NIST AI Risk Management Framework are treated as design inputs, so the evidence exists from the first release, in whichever jurisdictions you operate.
Yours to own
The architecture does not depend on any single model vendor, cloud or platform. You own the design, the code and the record of why each decision was made. Models sit behind a common interface, so replacing one is a planned engineering change with its own tests. Your team understands the system well enough to run it, extend it and, if it ever needs to, move it to another supplier.
Accountable value
Every system has a cost and a return, stated in terms the business already uses: cost per case handled, hours returned to a team, errors avoided, revenue influenced. Running cost is visible per use case and per model, so you can see when a cheaper model would do the same job. Each use case has a target agreed before it is built, and results are reported on the same schedule as the rest of the business. When a use case does not pay, you see it early and can stop it.
What to ask your team this quarter
Starting does not require a new strategy. It requires clear answers to a few questions:
- Which of our AI systems can take actions, and where is it written down what each one is allowed to do?
- If an agent did something it should not, how quickly would we know, and could we show exactly what happened?
- If our main model vendor doubled its prices or withdrew a model, what would switching cost us, and how long would it take?
- Who holds the code, the designs and the knowledge for each system? Could our own team change it without the supplier?
- What did AI cost us last quarter, in total and per use case, and what did each use case return?
- In every market we operate in, could we show a regulator today how a given system is controlled?
Where the answers are vague, that is the place to start. In most organisations, the weak answers will point to the architecture.
What we do
TaoQ AI is an independent AI architecture and engineering firm. We are not tied to any model vendor, cloud or platform. We design and build AI systems that stay under your control, work with any model, and show what they cost and deliver. Our work runs in three stages: a review of what you have, design and build of the architecture and the systems on it, and running and evolving them with your team until it can carry on without us. The way in is a fixed-fee review.